Opus Magnum Bench: leaderboard
Alchemy-machine construction on 36 campaign puzzles. The primary leaderboard takes the best solved reward per puzzle across published runs, then averages over all 36 puzzles. Reward is human-best cost plus cycles plus area divided by the solution total, capped at one; scores are percentages. Historical corrected trials follow current solved rewards, and evaluated zero-score models remain present. Run coverage differs between models; this is not pass@1.
Metric: Human-normalized score (%). Source: opusmagnumbench.com. Saturation forecast: Around January 2027. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (xHigh) | 70.86 | #19 (GPT-5.6 Sol) |
| 2 | Claude Fable 5 (High) | 60.15 | #16 (Claude Fable 5) |
| 3 | GPT-5.5 (xHigh) | 53.86 | #31 (GPT-5.5) |
| 4 | Claude Fable 5 (Low) | 49.28 | #16 (Claude Fable 5) |
| 5 | GPT-5.6 Terra (xHigh) | 47.72 | #41 (GPT-5.6 Terra) |
| 6 | GPT-5.5 (Medium) | 37.86 | #31 (GPT-5.5) |
| 7 | Gemini 3.1 Pro (Preview) (High) | 31.11 | #58 (Gemini 3.1 Pro (Preview)) |
| 8 | Claude Sonnet 5 (High) | 26.55 | #57 (Claude Sonnet 5) |
| 9 | GLM-5.2 (High) | 25.39 | #73 (GLM-5.2) |
| 10 | Gemini 3.5 Flash (High) | 24.49 | #60 (Gemini 3.5 Flash) |
| 11 | Claude Opus 4.8 (High) | 22.81 | #37 (Claude Opus 4.8) |
| 12 | Grok 4.5 (High) | 18.19 | #48 (Grok 4.5) |
| 13 | GPT-5.6 Luna (xHigh) | 17.8 | #68 (GPT-5.6 Luna) |
| 14 | Gemini 3 Flash (Preview) (High) | 15.39 | #82 (Gemini 3 Flash (Preview)) |
| 15 | DeepSeek V4 Pro | 13.39 | #102 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=opus-magnum-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.